Convolutional Neural Networks

Project: Write an Algorithm for a Dog Identification App


In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!

Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the Jupyter Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.

In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.

Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.

The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this Jupyter notebook.


Why We're Here

In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).

Sample Dog Output

In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!

The Road Ahead

We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.

  • Step 0: Import Datasets
  • Step 1: Detect Humans
  • Step 2: Detect Dogs
  • Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
  • Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)
  • Step 5: Write your Algorithm
  • Step 6: Test Your Algorithm

Step 0: Import Datasets

Make sure that you've downloaded the required human and dog datasets:

  • Download the dog dataset. Unzip the folder and place it in this project's home directory, at the location /dogImages.

  • Download the human dataset. Unzip the folder and place it in the home directory, at location /lfw.

Note: If you are using a Windows machine, you are encouraged to use 7zip to extract the folder.

In the code cell below, we save the file paths for both the human (LFW) dataset and dog dataset in the numpy arrays human_files and dog_files.

In [1]:
import numpy as np
from glob import glob

# load filenames for human and dog images
human_files = np.array(glob("lfw/*/*"))
dog_files = np.array(glob("dogImages/*/*/*"))

# print number of images in each dataset
print('There are %d total human images.' % len(human_files))
print('There are %d total dog images.' % len(dog_files))
There are 13233 total human images.
There are 8351 total dog images.

Step 1: Detect Humans

In this section, we use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images.

OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory. In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.

In [2]:
import cv2                
import matplotlib.pyplot as plt                        
%matplotlib inline                               

# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')

# load color (BGR) image
img = cv2.imread(human_files[0])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# find faces in image
faces = face_cascade.detectMultiScale(gray)

# print number of faces detected in the image
print('Number of faces detected:', len(faces))

# get blue bounding box for each detected face
for (x,y,w,h) in faces:
    # add bounding box to color image
    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
    
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
Number of faces detected: 1

Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.

In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.

Write a Human Face Detector

We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.

In [3]:
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
    img = cv2.imread(img_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray)
    return len(faces) > 0

Later in the project, I used a pre-trained PyTorch face detection model facenet-pytorch to identify whether a human is present

(IMPLEMENTATION) Assess the Human Face Detector

Question 1: Use the code cell below to test the performance of the face_detector function.

  • What percentage of the first 100 images in human_files have a detected human face?
  • What percentage of the first 100 images in dog_files have a detected human face?

Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.

Answer: (You can print out your results and/or write your percentages in this cell)

Correctly identified human faces: 96.0% \ Misclassified dog faces as human faces: 18.0%

In [4]:
from tqdm import tqdm

human_files_short = human_files[:100]
dog_files_short = dog_files[:100]

#-#-# Do NOT modify the code above this line. #-#-#

n_images = len(human_files_short)

print('Checking the true positive percentage in the first 100 images..')
correct_human = 0
for human in tqdm(human_files_short):
    
    # load color (BGR) image
    img = cv2.imread(human)
    
    # convert BGR image to grayscale
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    # find faces in image
    faces = face_cascade.detectMultiScale(gray)

    # print number of faces detected in the image
    correct_human += 1 if len(faces) > 0 else 0

    
print('Checking the false positive percentage in the first 100 images..')
incorrect_human = 0
for dog in tqdm(dog_files_short):
    
    # load color (BGR) image
    img = cv2.imread(dog)
    
    # convert BGR image to grayscale
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    # find faces in image
    faces = face_cascade.detectMultiScale(gray)

    # print number of faces detected in the image
    incorrect_human += 1 if len(faces) > 0 else 0
  4%|███▎                                                                              | 4/100 [00:00<00:02, 37.98it/s]
Checking the true positive percentage in the first 100 images..
100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [00:02<00:00, 36.58it/s]
  2%|█▋                                                                                | 2/100 [00:00<00:06, 16.13it/s]
Checking the false positive percentage in the first 100 images..
100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [00:12<00:00,  7.73it/s]
In [5]:
print(f'''
    Correctly identified human faces: {correct_human / n_images * 100}%
    Misclassified dog faces as human faces: {incorrect_human / n_images * 100}%
''')
    Correctly identified human faces: 96.0%
    Misclassified dog faces as human faces: 18.0%

We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.

---

Step 2: Detect Dogs

In this section, we use a pre-trained model to detect dogs in images.

Obtain Pre-trained VGG-16 Model

The code cell below downloads the VGG-16 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories.

In [6]:
import torch
import torchvision.models as models

# check if CUDA is available
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

Given an image, this pre-trained VGG-16 model returns a prediction (derived from the 1000 possible categories in ImageNet) for the object that is contained in the image.

(IMPLEMENTATION) Making Predictions with a Pre-trained Model

In the next code cell, you will write a function that accepts a path to an image (such as 'dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg') as input and returns the index corresponding to the ImageNet class that is predicted by the pre-trained VGG-16 model. The output should always be an integer between 0 and 999, inclusive.

Before writing the function, make sure that you take the time to learn how to appropriately pre-process tensors for pre-trained models in the PyTorch documentation.

In [7]:
from PIL import Image, ImageFile
import torchvision.transforms as transforms 
import requests

# Set PIL to be tolerant of image files that are truncated.
ImageFile.LOAD_TRUNCATED_IMAGES = True

def VGG16_predict(img_path: str) -> str:
    '''
    Use pre-trained VGG-16 model to obtain index corresponding to 
    predicted ImageNet class for image at specified path
    
    Args:
        img_path: path to an image
        
    Returns:
        Index corresponding to VGG-16 model's prediction
    '''
    
    # fetch imagenet class lables
    LABELS_URL = 'https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json'
    response = requests.get(LABELS_URL)  
    
    # create dictionary from the labels
    labels = {int(key): value[1] for key, value in response.json().items()}
    VGG16 = models.vgg16(pretrained=True)
    
    ## Load image
    img = Image.open(img_path).convert('RGB')
    
    transform = transforms.Compose([transforms.Resize(256), 
                                   transforms.CenterCrop(224), 
                                   transforms.ToTensor(),
                                   transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                                        std=[0.229, 0.224, 0.225])])
    
    # transform image
    img = transform(img)
    img = img[None, ...].float() # account for batch size
    
    # predict
    VGG16.eval() # ensure evaluation mode
    with torch.no_grad(): 
        output = VGG16(img)
     
    prediction = output.data.to('cpu').numpy().argmax() # return the index for largest value
    return prediction

VGG16_predict('dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg')
Out[7]:
252

(IMPLEMENTATION) Write a Dog Detector

While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained VGG-16 model, we need only check if the pre-trained model predicts an index between 151 and 268 (inclusive).

Use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).

In [8]:
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
    outputIndex = VGG16_predict(img_path)
    return True if outputIndex >= 151 and outputIndex <= 268 else False

(IMPLEMENTATION) Assess the Dog Detector

Question 2: Use the code cell below to test the performance of your dog_detector function.

  • What percentage of the images in human_files_short have a detected dog?
  • What percentage of the images in dog_files_short have a detected dog?

Answer:

Incorrect Dog 0.0% \ Correct Dog 94.0%

In [9]:
print('Checking the false positive percentage in the first 100 human images...')
incorrect_dog = 0
for human in tqdm(human_files_short):
    
    is_dog = dog_detector(human)
    incorrect_dog += 1 if is_dog else 0
    
print('Checking the true positive percentage in the first 100 dog images...')
correct_dog = 0
for dog in tqdm(dog_files_short):
    
    is_dog = dog_detector(dog)
    correct_dog += 1 if is_dog else 0
  0%|                                                                                          | 0/100 [00:00<?, ?it/s]
Checking the false positive percentage in the first 100 human images...
100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [02:49<00:00,  1.70s/it]
  0%|                                                                                          | 0/100 [00:00<?, ?it/s]
Checking the true positive percentage in the first 100 dog images...
100%|████████████████████████████████████████████████████████████████████████████████| 100/100 [02:52<00:00,  1.72s/it]
In [10]:
print(f'''
    Classified human as a dog {incorrect_dog / n_images * 100}%
    classified dog as a a dog {correct_dog / n_images * 100}%
''')
    Classified human as a dog 0.0%
    classified dog as a a dog 94.0%

We suggest VGG-16 as a potential network to detect dog images in your algorithm, but you are free to explore other pre-trained networks (such as Inception-v3, ResNet-50, etc). Please use the code cell below to test other pre-trained PyTorch models. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.


Step 3: Create a CNN to Classify Dog Breeds (from Scratch)

Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 10%. In Step 4 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.

We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have trouble distinguishing between a Brittany and a Welsh Springer Spaniel.

Brittany Welsh Springer Spaniel

It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).

Curly-Coated Retriever American Water Spaniel

Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.

Yellow Labrador Chocolate Labrador Black Labrador

We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.

Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!

(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset

Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dogImages/train, dogImages/valid, and dogImages/test, respectively). You may find this documentation on custom datasets to be a useful resource. If you are interested in augmenting your training and/or validation data, check out the wide variety of transforms!

In [11]:
import os
from torchvision import datasets

### TODO: Write data loaders for training, validation, and test sets
## Specify appropriate transforms, and batch_sizes

DATA_DIR = 'dogImages/'
TRAIN = 'train'
VAL = 'valid'
TEST = 'test'
KERNEL_SIZE = 3
SIGMA = (0.3, 1.8)
ROTATION = 30
RESIZE = (256, 256)
CROP = (224, 224)
BATCH_SIZE = 8

data_transforms = { 
    TRAIN: transforms.Compose([ # define train set augmentations
        transforms.Resize(RESIZE),
        transforms.RandomCrop(CROP),
        transforms.RandomVerticalFlip(),
        transforms.RandomRotation(ROTATION),
        transforms.RandomHorizontalFlip(),
        transforms.GaussianBlur(KERNEL_SIZE, sigma=SIGMA),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ]),
    VAL: transforms.Compose([ # no need to perform any augmentation on the validation data
        transforms.Resize(RESIZE),
        transforms.CenterCrop(CROP),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ]),
    TEST: transforms.Compose([ # no need to perform any augmentation on the test data
        transforms.Resize(RESIZE),
        transforms.CenterCrop(CROP),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406],
                             std=[0.229, 0.224, 0.225])
    ])
}

image_datasets = {
    x: datasets.ImageFolder(
        os.path.join(DATA_DIR, x), 
        transform=data_transforms[x]
    )
    for x in [TRAIN, VAL, TEST]
}

loaders_scratch = {
    x: torch.utils.data.DataLoader(
        image_datasets[x], batch_size=BATCH_SIZE,
        shuffle=True, num_workers=0  # turn on shuffle (though not needed for testing and validation)
    ) 
    for x in [TRAIN, VAL, TEST]
}

for dataset in image_datasets:
    n_images = len(image_datasets[dataset])
    print(f'Dataset {dataset} contains {n_images} images')
Dataset train contains 6680 images
Dataset valid contains 835 images
Dataset test contains 836 images

Question 3: Describe your chosen procedure for preprocessing the data.

  • How does your code resize the images (by cropping, stretching, etc)? What size did you pick for the input tensor, and why?
  • Did you decide to augment the dataset? If so, how (through translations, flips, rotations, etc)? If not, why not?

Answer:

  • The VGG model expects a 224 by 224 pixel images. \ Online, I noticed it's considered a good practice to resize the images to 256 by 256, \ then cropping the center 224 by 224 area. Additionally, the VGG archticture expects the pixel values \ to be normalized differently each color channel. mean = [0.485, 0.456, 0.406], and std=[0.229, 0.224, 0.225]

    Since the image domain I'm using is similar to ImageNet, I'm using the same mean and standard deviation \ to normalize pixel values. Also, I decided to also use the 224 by 224 size as it's small enough to allow \ me to train a batchs of images on my NVIDIA GPU that has 4 gigabytes memorry, but large enough to \ maintain details from the original image

  • The training set contains 6680 images. That's a small number of images considering the number of classes. \ Since the model is likely to overfit before achieving an acceptable classification accuracy, I decided to perform image augmentation.

    I randomly crop a 224 by 224 pixel area of the picture which would likely move what I want to classify in the images around (translation invariance) \ Also, I rotate and flip the images (rotation invariance). Finally, I'm experementing with Gaussian Blur. \ Since not everyone is great at taking pictures I'm expecting low quality user images. \ Hopefully, this will help make the model more robust to blurry images.

(IMPLEMENTATION) Model Architecture

Create a CNN to classify dog breed. Use the template in the code cell below.

In [12]:
import torch.nn as nn
import torch.nn.functional as F

# define the CNN architecture
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        
        ## define convolutional blocks
        self.convBlock1 = nn.Sequential(
            nn.Conv2d(3, 16, 3, padding=1),
            nn.Conv2d(16, 16, 3, padding=1),
            nn.MaxPool2d(2),
            nn.ReLU()
        )
        
        self.convBlock2 = nn.Sequential(
            nn.Conv2d(16, 32, 3, padding=1),
            nn.MaxPool2d(2),
            nn.ReLU()
        )
        
        self.convBlock3 = nn.Sequential(
            nn.Conv2d(32, 64, 3, padding=1),
            nn.MaxPool2d(2),
            nn.ReLU()
        )
        
        self.convBlock4 = nn.Sequential(
            nn.Conv2d(64, 128, 3, padding=1),
            nn.MaxPool2d(2),
            nn.ReLU()
        )
        
        ## define linear block
        self.linearBlock = nn.Sequential(
            nn.Linear(128 * 14 * 14, 516),
            nn.ReLU(), 
            nn.Dropout(0.3),
            nn.Linear(516, 256),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(256, 133),
        )
        pass
        
    def forward(self, x):
    
        x = self.convBlock1(x) # first convolution block
        x = self.convBlock2(x) # second convolution block
        x = self.convBlock3(x) # third convolution block
        x = self.convBlock4(x) # fourth convolution block
        x = x.view(x.size(0), -1) # flatten
        x = self.linearBlock(x) # pass through the linear layersThe b
        return x

#-#-# You do NOT have to modify the code below this line. #-#-#

# instantiate the CNN
model_scratch = Net()

Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step.

Answer:

The basic VGG16 structure contains multiple convolution blocks containing two convolution layers followed by max-pooling. I tried to maintain a similar architecture within reason.

Keeping in mind the number of trainable parameters, each convolution block in my architecture consists of a single convolution layer and max-pooling. Also, I reduced the model depth by cutting the output channels in half in each convolution layer. Finally, I removed the fifth convolution block.

Lastly, I prepare the input for the linear portion of the model. Again, similar to VGG16, I chose to include three linear layers and dropout regularization after every hidden layer output. However, the number of neurons in every layer is significantly smaller. The original dataset contains a thousand classes, while there are only 133 dog breeds in the dataset.

I chose the ReLU activation function on every hidden layer output. In most cases, ReLU returns the best results and is currently the go-to activation function. As this is a multi-class classification problem, I decided to use CrossEntropyLoss, which applies log softmax and NLLLoss on the raw logits. I chose Adam as the optimizer with a learning rate of 1e-4. I'm not sure if Adam is the best optimizer for this task. I will try SGD next time.

(IMPLEMENTATION) Specify Loss Function and Optimizer

Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_scratch, and the optimizer as optimizer_scratch below.

In [13]:
import torch.optim as optim

### TODO: select loss function
criterion_scratch = nn.CrossEntropyLoss()

### TODO: select optimizer
optimizer_scratch = optim.Adam(model_scratch.parameters(), 1e-4)

(IMPLEMENTATION) Train and Validate the Model

Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_scratch.pt'.

In [14]:
# the following import is required for training to be robust to truncated images
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import sys

def train(n_epochs, loaders, model, optimizer, criterion, device, save_path):
    """returns trained model"""
    # initialize tracker for minimum validation loss
    valid_loss_min = np.Inf 
    train_batches = int(np.ceil(len(loaders['train'])))
    model = model.to(device)
    
    for epoch in range(1, n_epochs+1):
        # initialize variables to monitor training and validation loss
        total = 0
        train_loss = 0.0
        valid_loss = 0.0
        correct_acc = 0.0
        correct_valid = 0.0
        ###################
        # train the model #
        ###################
        model.train()
        for batch_idx, (data, target) in enumerate(loaders['train']):
            # move to GPU
            data, target = data.to(device), target.to(device)
            
            # reset optimizer every iteration
            optimizer.zero_grad()
            
            output = model(data)
            loss = criterion(output, target)
            loss.backward()
            optimizer.step()
            
            _, pred = torch.max(output, 1)
           
            equals = pred == target
            correct_acc += torch.sum(equals.type(torch.cuda.FloatTensor)).item()
            train_loss = train_loss + ((1 / (batch_idx + 1)) * (loss.item() - train_loss))
            total += target.size(0)
            
            sys.stdout.write('\r')
            sys.stdout.write(f"Epoch: {epoch}\tTraining batch {batch_idx} out of {train_batches}\tClassified correctly: {round(correct_acc / total, 5)}")
            sys.stdout.flush()
    
            ## find the loss and update the model parameters accordingly
            ## record the average training loss, using something like
            ## train_loss = train_loss + ((1 / (batch_idx + 1)) * (loss.data - train_loss))
            
        ######################    
        # validate the model #
        ######################

        
        with torch.no_grad():
            model.eval()
            for batch_idx, (data, target) in enumerate(loaders['valid']):
                
                # move to GPU
                data, target = data.to(device), target.to(device)
                
                ## update the average validation loss
                output = model(data)
                loss = criterion(output, target)
                
                _, pred = torch.max(output, 1)
                equals = pred == target
                correct_valid += torch.sum(equals.type(torch.cuda.FloatTensor)).item()
                valid_loss = train_loss + ((1 / (batch_idx + 1)) * (loss.item() - train_loss))
                
                
        # print training/validation statistics 
        print('\nEpoch: {} \tTraining Loss: {:.6f} \tValidation Loss: {:.6f} \t Training accuracy: {:.6f}'.format(
            epoch, 
            train_loss,
            valid_loss,
            correct_acc / total
            ))
        
        ## TODO: save the model if validation loss has decreased
        if valid_loss < valid_loss_min:
            torch.save(model, save_path)
            print(f'\nValidation loss improved from {valid_loss_min} to {valid_loss}')
            valid_loss_min = valid_loss
        else: 
            print("Loss hasn't improved. Model not saved\n")
    
    model = model.to(torch.device('cpu'))
    return 

# train the model
train(80, loaders_scratch, model_scratch, optimizer_scratch, 
                  criterion_scratch, device, 'model_scratch.pt')
Epoch: 1	Training batch 834 out of 835	Classified correctly: 0.00749
Epoch: 1 	Training Loss: 4.880158 	Validation Loss: 4.880171 	 Training accuracy: 0.007485

Validation loss improved from inf to 4.880170542219614
Epoch: 2	Training batch 834 out of 835	Classified correctly: 0.02126
Epoch: 2 	Training Loss: 4.725266 	Validation Loss: 4.720577 	 Training accuracy: 0.021257

Validation loss improved from 4.880170542219614 to 4.720577110152383
Epoch: 3	Training batch 834 out of 835	Classified correctly: 0.03204
Epoch: 3 	Training Loss: 4.523835 	Validation Loss: 4.521556 	 Training accuracy: 0.032036

Validation loss improved from 4.720577110152383 to 4.5215555895078525
Epoch: 4	Training batch 834 out of 835	Classified correctly: 0.03548
Epoch: 4 	Training Loss: 4.415727 	Validation Loss: 4.412495 	 Training accuracy: 0.035479

Validation loss improved from 4.5215555895078525 to 4.412495258979583
Epoch: 5	Training batch 834 out of 835	Classified correctly: 0.04102
Epoch: 5 	Training Loss: 4.347043 	Validation Loss: 4.344929 	 Training accuracy: 0.041018

Validation loss improved from 4.412495258979583 to 4.34492915273426
Epoch: 6	Training batch 834 out of 835	Classified correctly: 0.04856
Epoch: 6 	Training Loss: 4.271404 	Validation Loss: 4.274061 	 Training accuracy: 0.048503

Validation loss improved from 4.34492915273426 to 4.274061346698561
Epoch: 7	Training batch 834 out of 835	Classified correctly: 0.05105
Epoch: 7 	Training Loss: 4.215513 	Validation Loss: 4.216394 	 Training accuracy: 0.051048

Validation loss improved from 4.274061346698561 to 4.216393893888818
Epoch: 8	Training batch 834 out of 835	Classified correctly: 0.06213
Epoch: 8 	Training Loss: 4.147255 	Validation Loss: 4.157440 	 Training accuracy: 0.062126

Validation loss improved from 4.216393893888818 to 4.157439768100214
Epoch: 9	Training batch 834 out of 835	Classified correctly: 0.06437
Epoch: 9 	Training Loss: 4.100965 	Validation Loss: 4.095351 	 Training accuracy: 0.064371

Validation loss improved from 4.157439768100214 to 4.095350880860805
Epoch: 10	Training batch 834 out of 835	Classified correctly: 0.07269
Epoch: 10 	Training Loss: 4.035567 	Validation Loss: 4.036155 	 Training accuracy: 0.072605

Validation loss improved from 4.095350880860805 to 4.036155173549429
Epoch: 11	Training batch 834 out of 835	Classified correctly: 0.07769
Epoch: 11 	Training Loss: 4.017524 	Validation Loss: 4.010395 	 Training accuracy: 0.077695

Validation loss improved from 4.036155173549429 to 4.010395290506924
Epoch: 12	Training batch 834 out of 835	Classified correctly: 0.07665
Epoch: 12 	Training Loss: 3.957930 	Validation Loss: 3.959855 	 Training accuracy: 0.076647

Validation loss improved from 4.010395290506924 to 3.9598546638684575
Epoch: 13	Training batch 834 out of 835	Classified correctly: 0.08398
Epoch: 13 	Training Loss: 3.917724 	Validation Loss: 3.914635 	 Training accuracy: 0.083982

Validation loss improved from 3.9598546638684575 to 3.9146354994299335
Epoch: 14	Training batch 834 out of 835	Classified correctly: 0.08937
Epoch: 14 	Training Loss: 3.884907 	Validation Loss: 3.878871 	 Training accuracy: 0.089371

Validation loss improved from 3.9146354994299335 to 3.8788707492845758
Epoch: 15	Training batch 834 out of 835	Classified correctly: 0.09042
Epoch: 15 	Training Loss: 3.838957 	Validation Loss: 3.843939 	 Training accuracy: 0.090419

Validation loss improved from 3.8788707492845758 to 3.8439385372789485
Epoch: 16	Training batch 834 out of 835	Classified correctly: 0.09491
Epoch: 16 	Training Loss: 3.808263 	Validation Loss: 3.805830 	 Training accuracy: 0.094910

Validation loss improved from 3.8439385372789485 to 3.805829882960323
Epoch: 17	Training batch 834 out of 835	Classified correctly: 0.11123
Epoch: 17 	Training Loss: 3.752594 	Validation Loss: 3.748701 	 Training accuracy: 0.111228

Validation loss improved from 3.805829882960323 to 3.748700667575176
Epoch: 18	Training batch 834 out of 835	Classified correctly: 0.10958
Epoch: 18 	Training Loss: 3.727321 	Validation Loss: 3.725678 	 Training accuracy: 0.109581

Validation loss improved from 3.748700667575176 to 3.7256781781199058
Epoch: 19	Training batch 834 out of 835	Classified correctly: 0.11497
Epoch: 19 	Training Loss: 3.686273 	Validation Loss: 3.693725 	 Training accuracy: 0.114970

Validation loss improved from 3.7256781781199058 to 3.6937254725680013
Epoch: 20	Training batch 834 out of 835	Classified correctly: 0.12036
Epoch: 20 	Training Loss: 3.661240 	Validation Loss: 3.658635 	 Training accuracy: 0.120359

Validation loss improved from 3.6937254725680013 to 3.658634593505143
Epoch: 21	Training batch 834 out of 835	Classified correctly: 0.11766
Epoch: 21 	Training Loss: 3.637233 	Validation Loss: 3.633851 	 Training accuracy: 0.117665

Validation loss improved from 3.658634593505143 to 3.633850865444292
Epoch: 22	Training batch 834 out of 835	Classified correctly: 0.12695
Epoch: 22 	Training Loss: 3.603928 	Validation Loss: 3.624064 	 Training accuracy: 0.126946

Validation loss improved from 3.633850865444292 to 3.6240640825663726
Epoch: 23	Training batch 834 out of 835	Classified correctly: 0.13263
Epoch: 23 	Training Loss: 3.559602 	Validation Loss: 3.560956 	 Training accuracy: 0.132635

Validation loss improved from 3.6240640825663726 to 3.5609561228391793
Epoch: 24	Training batch 834 out of 835	Classified correctly: 0.13353
Epoch: 24 	Training Loss: 3.526953 	Validation Loss: 3.531970 	 Training accuracy: 0.133533

Validation loss improved from 3.5609561228391793 to 3.531970023004835
Epoch: 25	Training batch 834 out of 835	Classified correctly: 0.14296
Epoch: 25 	Training Loss: 3.501149 	Validation Loss: 3.505443 	 Training accuracy: 0.142964

Validation loss improved from 3.531970023004835 to 3.505443072722853
Epoch: 26	Training batch 834 out of 835	Classified correctly: 0.14117
Epoch: 26 	Training Loss: 3.489015 	Validation Loss: 3.503274 	 Training accuracy: 0.141168

Validation loss improved from 3.505443072722853 to 3.5032740668064823
Epoch: 27	Training batch 834 out of 835	Classified correctly: 0.14793
Epoch: 27 	Training Loss: 3.457975 	Validation Loss: 3.475880 	 Training accuracy: 0.147904

Validation loss improved from 3.5032740668064823 to 3.4758803950587893
Epoch: 28	Training batch 834 out of 835	Classified correctly: 0.14671
Epoch: 28 	Training Loss: 3.410536 	Validation Loss: 3.431231 	 Training accuracy: 0.146707

Validation loss improved from 3.4758803950587893 to 3.4312308759066577
Epoch: 29	Training batch 834 out of 835	Classified correctly: 0.15763
Epoch: 29 	Training Loss: 3.398565 	Validation Loss: 3.406000 	 Training accuracy: 0.157635

Validation loss improved from 3.4312308759066577 to 3.4059999251658013
Epoch: 30	Training batch 834 out of 835	Classified correctly: 0.16183
Epoch: 30 	Training Loss: 3.373692 	Validation Loss: 3.385100 	 Training accuracy: 0.161826

Validation loss improved from 3.4059999251658013 to 3.385100401439729
Epoch: 31	Training batch 834 out of 835	Classified correctly: 0.16557
Epoch: 31 	Training Loss: 3.357967 	Validation Loss: 3.356200 	 Training accuracy: 0.165569

Validation loss improved from 3.385100401439729 to 3.356200041853874
Epoch: 32	Training batch 834 out of 835	Classified correctly: 0.16916
Epoch: 32 	Training Loss: 3.333135 	Validation Loss: 3.343171 	 Training accuracy: 0.169162

Validation loss improved from 3.356200041853874 to 3.34317058279333
Epoch: 33	Training batch 834 out of 835	Classified correctly: 0.17591
Epoch: 33 	Training Loss: 3.295854 	Validation Loss: 3.282064 	 Training accuracy: 0.175000

Validation loss improved from 3.34317058279333 to 3.2820644517171953
Epoch: 34	Training batch 834 out of 835	Classified correctly: 0.17829
Epoch: 34 	Training Loss: 3.281569 	Validation Loss: 3.285008 	 Training accuracy: 0.178293
Loss hasn't improved. Model not saved

Epoch: 35	Training batch 834 out of 835	Classified correctly: 0.18204
Epoch: 35 	Training Loss: 3.252766 	Validation Loss: 3.268096 	 Training accuracy: 0.182036

Validation loss improved from 3.2820644517171953 to 3.2680959415952446
Epoch: 36	Training batch 834 out of 835	Classified correctly: 0.18009
Epoch: 36 	Training Loss: 3.227541 	Validation Loss: 3.234200 	 Training accuracy: 0.180090

Validation loss improved from 3.2680959415952446 to 3.2341995073840084
Epoch: 37	Training batch 834 out of 835	Classified correctly: 0.19251
Epoch: 37 	Training Loss: 3.212471 	Validation Loss: 3.214199 	 Training accuracy: 0.192515

Validation loss improved from 3.2341995073840084 to 3.2141990748558533
Epoch: 38	Training batch 834 out of 835	Classified correctly: 0.19716
Epoch: 38 	Training Loss: 3.185717 	Validation Loss: 3.178651 	 Training accuracy: 0.197156

Validation loss improved from 3.2141990748558533 to 3.1786512182239517
Epoch: 39	Training batch 834 out of 835	Classified correctly: 0.19431
Epoch: 39 	Training Loss: 3.155716 	Validation Loss: 3.161627 	 Training accuracy: 0.194311

Validation loss improved from 3.1786512182239517 to 3.161627010536358
Epoch: 40	Training batch 834 out of 835	Classified correctly: 0.19985
Epoch: 40 	Training Loss: 3.146754 	Validation Loss: 3.137669 	 Training accuracy: 0.199850

Validation loss improved from 3.161627010536358 to 3.1376693921460364
Epoch: 41	Training batch 834 out of 835	Classified correctly: 0.20868
Epoch: 41 	Training Loss: 3.116212 	Validation Loss: 3.115617 	 Training accuracy: 0.208683

Validation loss improved from 3.1376693921460364 to 3.1156167952709133
Epoch: 42	Training batch 834 out of 835	Classified correctly: 0.20629
Epoch: 42 	Training Loss: 3.095233 	Validation Loss: 3.078091 	 Training accuracy: 0.206287

Validation loss improved from 3.1156167952709133 to 3.078090646470209
Epoch: 43	Training batch 834 out of 835	Classified correctly: 0.21722
Epoch: 43 	Training Loss: 3.065000 	Validation Loss: 3.066730 	 Training accuracy: 0.217216

Validation loss improved from 3.078090646470209 to 3.066729546604855
Epoch: 44	Training batch 834 out of 835	Classified correctly: 0.22006
Epoch: 44 	Training Loss: 3.058062 	Validation Loss: 3.070130 	 Training accuracy: 0.220060
Loss hasn't improved. Model not saved

Epoch: 45	Training batch 834 out of 835	Classified correctly: 0.23129
Epoch: 45 	Training Loss: 3.010194 	Validation Loss: 3.006597 	 Training accuracy: 0.231287

Validation loss improved from 3.066729546604855 to 3.0065965143839506
Epoch: 46	Training batch 834 out of 835	Classified correctly: 0.22267
Epoch: 46 	Training Loss: 3.020492 	Validation Loss: 3.032778 	 Training accuracy: 0.222605
Loss hasn't improved. Model not saved

Epoch: 47	Training batch 834 out of 835	Classified correctly: 0.22575
Epoch: 47 	Training Loss: 2.995640 	Validation Loss: 2.980925 	 Training accuracy: 0.225749

Validation loss improved from 3.0065965143839506 to 2.9809249698955966
Epoch: 48	Training batch 834 out of 835	Classified correctly: 0.23278
Epoch: 48 	Training Loss: 2.976319 	Validation Loss: 2.978186 	 Training accuracy: 0.232784

Validation loss improved from 2.9809249698955966 to 2.97818642039771
Epoch: 49	Training batch 834 out of 835	Classified correctly: 0.23653
Epoch: 49 	Training Loss: 2.932175 	Validation Loss: 2.939028 	 Training accuracy: 0.236527

Validation loss improved from 2.97818642039771 to 2.939027823216222
Epoch: 50	Training batch 834 out of 835	Classified correctly: 0.23653
Epoch: 50 	Training Loss: 2.924211 	Validation Loss: 2.922970 	 Training accuracy: 0.236527

Validation loss improved from 2.939027823216222 to 2.922970110463863
Epoch: 51	Training batch 834 out of 835	Classified correctly: 0.24072
Epoch: 51 	Training Loss: 2.924519 	Validation Loss: 2.927437 	 Training accuracy: 0.240719
Loss hasn't improved. Model not saved

Epoch: 52	Training batch 834 out of 835	Classified correctly: 0.24626
Epoch: 52 	Training Loss: 2.884853 	Validation Loss: 2.894535 	 Training accuracy: 0.246257

Validation loss improved from 2.922970110463863 to 2.8945352098232697
Epoch: 53	Training batch 834 out of 835	Classified correctly: 0.25629
Epoch: 53 	Training Loss: 2.859479 	Validation Loss: 2.844891 	 Training accuracy: 0.256287

Validation loss improved from 2.8945352098232697 to 2.8448912863150784
Epoch: 54	Training batch 834 out of 835	Classified correctly: 0.26018
Epoch: 54 	Training Loss: 2.841410 	Validation Loss: 2.844259 	 Training accuracy: 0.260180

Validation loss improved from 2.8448912863150784 to 2.8442591569197297
Epoch: 55	Training batch 834 out of 835	Classified correctly: 0.25808
Epoch: 55 	Training Loss: 2.817436 	Validation Loss: 2.829439 	 Training accuracy: 0.258084

Validation loss improved from 2.8442591569197297 to 2.8294389533126383
Epoch: 56	Training batch 834 out of 835	Classified correctly: 0.26078
Epoch: 56 	Training Loss: 2.803979 	Validation Loss: 2.809579 	 Training accuracy: 0.260778

Validation loss improved from 2.8294389533126383 to 2.8095785381952387
Epoch: 57	Training batch 834 out of 835	Classified correctly: 0.26362
Epoch: 57 	Training Loss: 2.781974 	Validation Loss: 2.785076 	 Training accuracy: 0.263623

Validation loss improved from 2.8095785381952387 to 2.7850763005560424
Epoch: 58	Training batch 834 out of 835	Classified correctly: 0.26302
Epoch: 58 	Training Loss: 2.768349 	Validation Loss: 2.765782 	 Training accuracy: 0.263024

Validation loss improved from 2.7850763005560424 to 2.7657821673465857
Epoch: 59	Training batch 834 out of 835	Classified correctly: 0.26183
Epoch: 59 	Training Loss: 2.771019 	Validation Loss: 2.776155 	 Training accuracy: 0.261826
Loss hasn't improved. Model not saved

Epoch: 60	Training batch 834 out of 835	Classified correctly: 0.27713
Epoch: 60 	Training Loss: 2.736569 	Validation Loss: 2.746862 	 Training accuracy: 0.277096

Validation loss improved from 2.7657821673465857 to 2.746862286836087
Epoch: 61	Training batch 834 out of 835	Classified correctly: 0.27653
Epoch: 61 	Training Loss: 2.705127 	Validation Loss: 2.716855 	 Training accuracy: 0.276497

Validation loss improved from 2.746862286836087 to 2.7168547702073598
Epoch: 62	Training batch 834 out of 835	Classified correctly: 0.28054
Epoch: 62 	Training Loss: 2.705293 	Validation Loss: 2.708010 	 Training accuracy: 0.280539

Validation loss improved from 2.7168547702073598 to 2.708009839926751
Epoch: 63	Training batch 834 out of 835	Classified correctly: 0.29222
Epoch: 63 	Training Loss: 2.690710 	Validation Loss: 2.715583 	 Training accuracy: 0.292216
Loss hasn't improved. Model not saved

Epoch: 64	Training batch 834 out of 835	Classified correctly: 0.29401
Epoch: 64 	Training Loss: 2.650002 	Validation Loss: 2.656521 	 Training accuracy: 0.294012

Validation loss improved from 2.708009839926751 to 2.6565208123639175
Epoch: 65	Training batch 834 out of 835	Classified correctly: 0.29027
Epoch: 65 	Training Loss: 2.656162 	Validation Loss: 2.685395 	 Training accuracy: 0.290269
Loss hasn't improved. Model not saved

Epoch: 66	Training batch 834 out of 835	Classified correctly: 0.29835
Epoch: 66 	Training Loss: 2.613559 	Validation Loss: 2.600883 	 Training accuracy: 0.298353

Validation loss improved from 2.6565208123639175 to 2.600883107564709
Epoch: 67	Training batch 834 out of 835	Classified correctly: 0.30359
Epoch: 67 	Training Loss: 2.604068 	Validation Loss: 2.613826 	 Training accuracy: 0.303593
Loss hasn't improved. Model not saved

Epoch: 68	Training batch 834 out of 835	Classified correctly: 0.29826
Epoch: 68 	Training Loss: 2.594665 	Validation Loss: 2.601738 	 Training accuracy: 0.298204
Loss hasn't improved. Model not saved

Epoch: 69	Training batch 834 out of 835	Classified correctly: 0.30284
Epoch: 69 	Training Loss: 2.593040 	Validation Loss: 2.611753 	 Training accuracy: 0.302844
Loss hasn't improved. Model not saved

Epoch: 70	Training batch 834 out of 835	Classified correctly: 0.31407
Epoch: 70 	Training Loss: 2.561929 	Validation Loss: 2.579454 	 Training accuracy: 0.314072

Validation loss improved from 2.600883107564709 to 2.5794540118926412
Epoch: 71	Training batch 834 out of 835	Classified correctly: 0.31257
Epoch: 71 	Training Loss: 2.537316 	Validation Loss: 2.543343 	 Training accuracy: 0.312575

Validation loss improved from 2.5794540118926412 to 2.5433425582111284
Epoch: 72	Training batch 834 out of 835	Classified correctly: 0.32231
Epoch: 72 	Training Loss: 2.518479 	Validation Loss: 2.511646 	 Training accuracy: 0.322305

Validation loss improved from 2.5433425582111284 to 2.511645690824424
Epoch: 73	Training batch 834 out of 835	Classified correctly: 0.31961
Epoch: 73 	Training Loss: 2.508022 	Validation Loss: 2.542112 	 Training accuracy: 0.319611
Loss hasn't improved. Model not saved

Epoch: 74	Training batch 834 out of 835	Classified correctly: 0.32515
Epoch: 74 	Training Loss: 2.486054 	Validation Loss: 2.475159 	 Training accuracy: 0.325150

Validation loss improved from 2.511645690824424 to 2.4751588487326015
Epoch: 75	Training batch 834 out of 835	Classified correctly: 0.32539
Epoch: 75 	Training Loss: 2.475389 	Validation Loss: 2.481231 	 Training accuracy: 0.325299
Loss hasn't improved. Model not saved

Epoch: 76	Training batch 834 out of 835	Classified correctly: 0.33338
Epoch: 76 	Training Loss: 2.442711 	Validation Loss: 2.448000 	 Training accuracy: 0.333383

Validation loss improved from 2.4751588487326015 to 2.4479999933631698
Epoch: 77	Training batch 834 out of 835	Classified correctly: 0.33488
Epoch: 77 	Training Loss: 2.438993 	Validation Loss: 2.451142 	 Training accuracy: 0.334880
Loss hasn't improved. Model not saved

Epoch: 78	Training batch 834 out of 835	Classified correctly: 0.33728
Epoch: 78 	Training Loss: 2.442334 	Validation Loss: 2.428465 	 Training accuracy: 0.337275

Validation loss improved from 2.4479999933631698 to 2.4284648822548385
Epoch: 79	Training batch 834 out of 835	Classified correctly: 0.34536
Epoch: 79 	Training Loss: 2.391698 	Validation Loss: 2.400375 	 Training accuracy: 0.345359

Validation loss improved from 2.4284648822548385 to 2.4003754200974785
Epoch: 80	Training batch 834 out of 835	Classified correctly: 0.34857
Epoch: 80 	Training Loss: 2.392743 	Validation Loss: 2.410054 	 Training accuracy: 0.348503
Loss hasn't improved. Model not saved

(IMPLEMENTATION) Test the Model

Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 10%.

In [15]:
# load the model that got the best validation accuracy
model_scratch = torch.load('model_scratch.pt')

def test(loaders, model, criterion, device):

    # monitor test loss and accuracy
    test_loss = 0.
    correct = 0.
    total = 0.
    model = model.to(device)
    
    model.eval()
    for batch_idx, (data, target) in enumerate(loaders['test']):
        
        # move to gpu
        data, target = data.to(device), target.to(device)
        
        
        # forward pass: compute predicted outputs by passing inputs to the model
        output = model(data)
        # calculate the loss
        loss = criterion(output, target)
        # update average test loss 
        test_loss = test_loss + ((1 / (batch_idx + 1)) * (loss.data - test_loss))
        # convert output probabilities to predicted class
        pred = output.data.max(1, keepdim=True)[1]
        # compare predictions to true label
        correct += np.sum(np.squeeze(pred.eq(target.data.view_as(pred))).cpu().numpy())
        total += data.size(0)
            
    print('Test Loss: {:.6f}\n'.format(test_loss))
    print('\nTest Accuracy: %2d%% (%2d/%2d)' % (
        100. * correct / total, correct, total))
    
    # make sure to free up gpu memory
    data, target = data.to(torch.device("cpu")), target.to(torch.device("cpu"))
    model = model.to(torch.device("cpu"))
    pass


# call test function    
test(loaders_scratch, model_scratch, criterion_scratch, device)
Test Loss: 3.425704


Test Accuracy: 21% (179/836)

Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)

You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.

(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset

Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dogImages/train, dogImages/valid, and dogImages/test, respectively).

If you like, you are welcome to use the same data loaders from the previous step, when you created a CNN from scratch.

(IMPLEMENTATION) Model Architecture

Use transfer learning to create a CNN to classify dog breed. Use the code cell below, and save your initialized model as the variable model_transfer.

In [16]:
import torchvision.models as models
import torch.nn as nn

# use VGG16 
model_transfer = models.vgg16(pretrained=True)

# freeze layers
for param in model_transfer.features.parameters():
    param.require_grad = False
    
# remove last layer
classifier_block = model_transfer.classifier

num_features = model_transfer.classifier[-1].in_features # save the number of in features in the last layer
classifier_block = list(classifier_block[:-1]) # remove last layer

# replace the layer with a new output with the number of classes
classifier_block.extend([nn.Linear(num_features, 133)])

# replace classifier layer
model_transfer.classifier = nn.Sequential(*classifier_block)

Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.

Answer:

I chose to continue working with VGG16 for the transfer learning exercise. Firstly, I froze all the layers to prevent the model from adjusting the weights. My only change to the model architecture was to remove the last fully connected layer and replace it with one that matches the number of classes (breeds) in the dog images.
VGG16 achieved 97.66% accuracy on ImageNet classification. Since the type of images I'm attempting to classify are similar to those found in ImageNet, I suspect that this architecture is suitable for the task.

(IMPLEMENTATION) Specify Loss Function and Optimizer

Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_transfer, and the optimizer as optimizer_transfer below.

In [17]:
criterion_transfer = nn.CrossEntropyLoss() # worked well before
optimizer_transfer = optim.SGD(model_transfer.parameters(), lr=0.001, momentum=0.9) # SGD might perform better than Adam

(IMPLEMENTATION) Train and Validate the Model

Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_transfer.pt'.

In [18]:
# train the model 
n_epochs = 30
model_transfer = train(n_epochs, loaders_scratch, model_transfer, 
                       optimizer_transfer, criterion_transfer, 
                       device, 'model_transfer.pt')
Epoch: 1	Training batch 834 out of 835	Classified correctly: 0.26991
Epoch: 1 	Training Loss: 3.031482 	Validation Loss: 3.021390 	 Training accuracy: 0.269910

Validation loss improved from inf to 3.0213903547033465
Epoch: 2	Training batch 834 out of 835	Classified correctly: 0.45359
Epoch: 2 	Training Loss: 1.975106 	Validation Loss: 1.975186 	 Training accuracy: 0.453593

Validation loss improved from 3.0213903547033465 to 1.97518589340794
Epoch: 3	Training batch 834 out of 835	Classified correctly: 0.54925
Epoch: 3 	Training Loss: 1.598222 	Validation Loss: 1.589205 	 Training accuracy: 0.549251

Validation loss improved from 1.97518589340794 to 1.5892054198098924
Epoch: 4	Training batch 834 out of 835	Classified correctly: 0.59311
Epoch: 4 	Training Loss: 1.373187 	Validation Loss: 1.366306 	 Training accuracy: 0.593114

Validation loss improved from 1.5892054198098924 to 1.3663055825491792
Epoch: 5	Training batch 834 out of 835	Classified correctly: 0.64371
Epoch: 5 	Training Loss: 1.181792 	Validation Loss: 1.177171 	 Training accuracy: 0.643713

Validation loss improved from 1.3663055825491792 to 1.1771713162265593
Epoch: 6	Training batch 834 out of 835	Classified correctly: 0.67021
Epoch: 6 	Training Loss: 1.089559 	Validation Loss: 1.081334 	 Training accuracy: 0.670210

Validation loss improved from 1.1771713162265593 to 1.0813343959706847
Epoch: 7	Training batch 834 out of 835	Classified correctly: 0.70419
Epoch: 7 	Training Loss: 0.990986 	Validation Loss: 0.986020 	 Training accuracy: 0.704192

Validation loss improved from 1.0813343959706847 to 0.9860195527880297
Epoch: 8	Training batch 834 out of 835	Classified correctly: 0.72066
Epoch: 8 	Training Loss: 0.899839 	Validation Loss: 0.891643 	 Training accuracy: 0.720659

Validation loss improved from 0.9860195527880297 to 0.8916427994104555
Epoch: 9	Training batch 834 out of 835	Classified correctly: 0.72657
Epoch: 9 	Training Loss: 0.859160 	Validation Loss: 0.854189 	 Training accuracy: 0.726497

Validation loss improved from 0.8916427994104555 to 0.8541891330081172
Epoch: 10	Training batch 834 out of 835	Classified correctly: 0.75299
Epoch: 10 	Training Loss: 0.778442 	Validation Loss: 0.784050 	 Training accuracy: 0.752994

Validation loss improved from 0.8541891330081172 to 0.7840504074936283
Epoch: 11	Training batch 834 out of 835	Classified correctly: 0.76437
Epoch: 11 	Training Loss: 0.761164 	Validation Loss: 0.753983 	 Training accuracy: 0.764371

Validation loss improved from 0.7840504074936283 to 0.7539833315111072
Epoch: 12	Training batch 834 out of 835	Classified correctly: 0.78293
Epoch: 12 	Training Loss: 0.678053 	Validation Loss: 0.671734 	 Training accuracy: 0.782934

Validation loss improved from 0.7539833315111072 to 0.6717336010043078
Epoch: 13	Training batch 834 out of 835	Classified correctly: 0.80314
Epoch: 13 	Training Loss: 0.619279 	Validation Loss: 0.613502 	 Training accuracy: 0.803144

Validation loss improved from 0.6717336010043078 to 0.6135019049531852
Epoch: 14	Training batch 834 out of 835	Classified correctly: 0.80156
Epoch: 14 	Training Loss: 0.619321 	Validation Loss: 0.614869 	 Training accuracy: 0.801497
Loss hasn't improved. Model not saved

Epoch: 15	Training batch 834 out of 835	Classified correctly: 0.81961
Epoch: 15 	Training Loss: 0.561976 	Validation Loss: 0.563083 	 Training accuracy: 0.819611

Validation loss improved from 0.6135019049531852 to 0.5630828921965538
Epoch: 16	Training batch 834 out of 835	Classified correctly: 0.82874
Epoch: 16 	Training Loss: 0.538101 	Validation Loss: 0.536395 	 Training accuracy: 0.828743

Validation loss improved from 0.5630828921965538 to 0.536394762769321
Epoch: 17	Training batch 834 out of 835	Classified correctly: 0.83578
Epoch: 17 	Training Loss: 0.505726 	Validation Loss: 0.503245 	 Training accuracy: 0.835778

Validation loss improved from 0.536394762769321 to 0.5032449724397856
Epoch: 18	Training batch 834 out of 835	Classified correctly: 0.83892
Epoch: 18 	Training Loss: 0.495880 	Validation Loss: 0.495474 	 Training accuracy: 0.838922

Validation loss improved from 0.5032449724397856 to 0.4954740294557471
Epoch: 19	Training batch 834 out of 835	Classified correctly: 0.84955
Epoch: 19 	Training Loss: 0.462084 	Validation Loss: 0.466698 	 Training accuracy: 0.849551

Validation loss improved from 0.4954740294557471 to 0.46669840228030285
Epoch: 20	Training batch 834 out of 835	Classified correctly: 0.86647
Epoch: 20 	Training Loss: 0.429496 	Validation Loss: 0.425482 	 Training accuracy: 0.866467

Validation loss improved from 0.46669840228030285 to 0.4254820196612257
Epoch: 21	Training batch 834 out of 835	Classified correctly: 0.84596
Epoch: 21 	Training Loss: 0.463870 	Validation Loss: 0.462757 	 Training accuracy: 0.845958
Loss hasn't improved. Model not saved

Epoch: 22	Training batch 834 out of 835	Classified correctly: 0.86871
Epoch: 22 	Training Loss: 0.408139 	Validation Loss: 0.410146 	 Training accuracy: 0.868713

Validation loss improved from 0.4254820196612257 to 0.41014624031715785
Epoch: 23	Training batch 834 out of 835	Classified correctly: 0.87859
Epoch: 23 	Training Loss: 0.376928 	Validation Loss: 0.389903 	 Training accuracy: 0.878593

Validation loss improved from 0.41014624031715785 to 0.3899032545095563
Epoch: 24	Training batch 834 out of 835	Classified correctly: 0.87445
Epoch: 24 	Training Loss: 0.385067 	Validation Loss: 0.381490 	 Training accuracy: 0.874401

Validation loss improved from 0.3899032545095563 to 0.38148979097998836
Epoch: 25	Training batch 834 out of 835	Classified correctly: 0.88668
Epoch: 25 	Training Loss: 0.353533 	Validation Loss: 0.360704 	 Training accuracy: 0.886677

Validation loss improved from 0.38148979097998836 to 0.36070395388769294
Epoch: 26	Training batch 834 out of 835	Classified correctly: 0.89326
Epoch: 26 	Training Loss: 0.342472 	Validation Loss: 0.352149 	 Training accuracy: 0.893263

Validation loss improved from 0.36070395388769294 to 0.3521492473380003
Epoch: 27	Training batch 834 out of 835	Classified correctly: 0.88533
Epoch: 27 	Training Loss: 0.348027 	Validation Loss: 0.355771 	 Training accuracy: 0.885329
Loss hasn't improved. Model not saved

Epoch: 28	Training batch 834 out of 835	Classified correctly: 0.90195
Epoch: 28 	Training Loss: 0.320965 	Validation Loss: 0.327925 	 Training accuracy: 0.901946

Validation loss improved from 0.3521492473380003 to 0.327925227392185
Epoch: 29	Training batch 834 out of 835	Classified correctly: 0.89853
Epoch: 29 	Training Loss: 0.314536 	Validation Loss: 0.311717 	 Training accuracy: 0.898503

Validation loss improved from 0.327925227392185 to 0.31171742129700736
Epoch: 30	Training batch 834 out of 835	Classified correctly: 0.90614
Epoch: 30 	Training Loss: 0.286398 	Validation Loss: 0.283785 	 Training accuracy: 0.906138

Validation loss improved from 0.31171742129700736 to 0.2837847417444288

(IMPLEMENTATION) Test the Model

Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 60%.

In [19]:
# load the model that got the best validation accuracy (uncomment the line below)
model_transfer = torch.load('model_transfer.pt')
test(loaders_scratch, model_transfer, criterion_transfer, device)
Test Loss: 1.135668


Test Accuracy: 74% (622/836)

(IMPLEMENTATION) Predict Dog Breed with the Model

Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan hound, etc) that is predicted by your model.

In [22]:
import joblib as jb

# list of class names by index, i.e. a name can be accessed like class_names[0]
class_names = [item[4:].replace("_", " ") for item in image_datasets['train'].classes]
class_names = {i: class_names[i] for i in range(133)}
jb.dump(class_names, 'class_names.pkl')
Out[22]:
['class_names.pkl']
In [23]:
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.

def predict_breed_transfer(img_path):
    # load the image and return the predicted breed
    
    # load classes and model
    class_dict = jb.load('class_names.pkl')
    model = torch.load('model_transfer.pt')
    
    # open and process image
    img = Image.open(img_path).convert('RGB')
    
    transform = transforms.Compose([transforms.Resize(256), 
                                   transforms.CenterCrop(224), 
                                   transforms.ToTensor(),
                                   transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                                        std=[0.229, 0.224, 0.225])])
    
    # transform image
    img = transform(img)
    img = img[None, ...].float() # account for batch size
    img, model = img.to('cpu'), model.to('cpu')
    output = model(img)
    pred = output.data.max(1, keepdim=True)[1].item()
    
    return class_dict[pred]

# predict 
predict_breed_transfer('dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg')
Out[23]:
'Affenpinscher'

Step 5: Write your Algorithm

Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,

  • if a dog is detected in the image, return the predicted breed.
  • if a human is detected in the image, return the resembling dog breed.
  • if neither is detected in the image, provide output that indicates an error.

You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 4 to predict dog breed.

Some sample output for our algorithm is provided below, but feel free to design your own user experience!

Sample Human Output

(IMPLEMENTATION) Write your Algorithm

In [30]:
from PIL import ImageFile, Image
import sys
import numpy as np
import joblib as jb
import matplotlib.pyplot as plt

import torch
import torchvision.models as models
import torchvision.transforms as transforms 
import torch.nn.functional as F
import torch.optim as optim
import torch.nn as nn
from facenet_pytorch import MTCNN 

def face_detector(image_path: str) -> float:
    '''
    using pretrained VGG_FACE2 model to detect faces
    '''
    # check if CUDA is available
    cuda = torch.cuda.is_available()

    # load image
    image = Image.open(image_path).convert('RGB') # ensure color image
    
    # create a face detection pipeline using MTCNN (suggested parameters):
    mtcnn = MTCNN(
        image_size=160, margin=0, min_face_size=20, 
        thresholds=[0.6, 0.7, 0.7], factor=0.709, post_process=True
    ) # model trained on image 160 pixel size images 
    
    _, proba = mtcnn(image, return_prob=True)
    try:
        return round(proba, 6) 
    except:
        return 0 

def dog_detector(image_path: str) -> bool:
    '''
    using pretrained VGG model to identify images containing dogs
    '''
    
    # load image
    image =  prepare_image(image_path)
    
    # load pretrained ImageNet model
    VGG16 = models.vgg16(pretrained=True)
    
    # check if CUDA is available
    cuda = torch.cuda.is_available()

    # move model to GPU if CUDA is available
    
    VGG16.eval()
    with torch.no_grad():
        if cuda:
            VGG16 = VGG16.cuda()
            image = image.cuda()
        
        outputs = VGG16(image)
    
    pred = outputs.data.to('cpu').numpy().argmax() # return the index for largest value
    is_dog = True if pred >= 151 and pred <= 268 else False # check whether the image contains a dog
    
    return is_dog

def prepare_image(image_path: str) -> torch.Tensor:
    '''
    using pytorch transforms to load and procces image for dog breed classification
    '''
    
    # load image
    image = Image.open(image_path).convert('RGB') # ensure color image
        
    # process image for vgg 16
    transform = transforms.Compose([transforms.Resize(256), 
                                   transforms.CenterCrop(224), 
                                   transforms.ToTensor(),
                                   transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                                        std=[0.229, 0.224, 0.225])])
    
    img = transform(image) # apply transforms   
    img = img[None, ...].float() # account for batch size
    
    # ensure prepare_image returns the correct data type and shape
    assert type(img) == torch.Tensor, 'The function prepare_image() does not return\
    the expected data type. It is suppose to convert a PIL image to Tensor'

    assert img.shape == torch.Size([1, 3, 224, 224]), 'The function prepare_image() does not return\
    the expected data type. Expected shape [1, 3, 224, 224] - [batch, channels, height, width]'
        
    return img 

def run_app(img_path: str):
    
    assert type(img_path) == str, 'Function requires a image path as a string'
    
    # check whether the image contains an image of a dog
    is_dog = dog_detector(img_path)
    
    # check whether the image contains a face
    face = True if face_detector(img_path) > 0.975 else  False
    
    try:
        # load dog breed classifier
        dog_breed_labels = jb.load('class_names.pkl')
        dog_breed_model = torch.load('model_transfer.pt')

        # prepare image for breed classification
        image_tensor = prepare_image(img_path)

         # classify the breed
        if is_dog or face:
            dog_breed_model.eval() # ensure evaluation mode
            with torch.no_grad(): 
                dog_breed_model = dog_breed_model.to(device)
                image_tensor = image_tensor.to(device)
                output = dog_breed_model(image_tensor)        

            # return probabilities
            softmax = nn.Softmax(dim=1) 
            probailities = softmax(output)

            # extract top breeds
            n_breeds = 2
            proba, ind = torch.topk(probailities, n_breeds) 
            proba, ind = proba.squeeze(), ind.squeeze() 

            top_dogs = {dog_breed_labels[ind[i].item()]: round(proba[i].item()*100, 2) for i in range(n_breeds)}
            # compose user-message
            breeds = list(top_dogs.keys())
            probas = list(top_dogs.values())

        # find if there's a dog in the image
        if is_dog:

            if probas[0] >= 65:
                message = f'The dog in the image looks like a pure-bread {breeds[0]}'

            else:
                message = f'''The dog in the image seems to be at least {probas[0]}% {breeds[0]}\
                \nand {probas[1]}% {breeds[1]}'''

        # check if there's dog but there is a face
        elif face: 
            message = f'''The person in the image resembles {probas[0]}% {breeds[0]}\
            \nand {probas[1]}% {breeds[1]}'''

        else:
            message = 'Unable to classify the dog breed. Please try a different image'
    except Exception as e:
        print(str(e))
        message = 'Unable to classify the dog breed. Please try a different image'
        
    return message

Step 6: Test Your Algorithm

In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?

(IMPLEMENTATION) Test Your Algorithm on Sample Images!

Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.

Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.

Answer: (Three possible points for improvement)

  1. Train the dog breed classifier longer to achieve more than 70%~ accuracy on the test set. I stopped after 25 epochs. Due to the augmentation, I believe the model could perform better before overfitting to the train set.
  2. Apply transfer learning to create a human classifier model instead of the pre-trained face detector that performs too well. For example, the face detector sometimes identifies dogs' faces with high probability, similar to human faces. That's why I chose 0.975 as the cutoff point to decide whether a face is human. Although humans are not part of ImageNet labels. It was said that the models detect humans as a feature).
  3. Instead of returning just the original image with the model outputs for humans, I could return the original image, the resembling dog, and a mash-up between the pictures laid out side by side..
  4. Visualize the intermediate model outputs and identify what parts of the images the model uses as features to identify each breed.
In [31]:
%matplotlib inline

## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.

for file in np.hstack((human_files[:3], dog_files[:3])):
    print(run_app(str(file)) + '\n')
The person in the image resembles 80.6% Dogue de bordeaux            
and 18.55% Pharaoh hound

The person in the image resembles 98.23% Dogue de bordeaux            
and 0.95% Pharaoh hound

The person in the image resembles 96.25% Dogue de bordeaux            
and 3.4% Chesapeake bay retriever

The dog in the image seems to be at least 64.8% Black russian terrier                
and 17.57% Briard

The dog in the image looks like a pure-bread Affenpinscher

The dog in the image looks like a pure-bread Brussels griffon

In [32]:
# random number for starting index
human_index = np.random.randint(100) 

# select images
test_humans = human_files[human_index:human_index+9] 

fig, ax = plt.subplots(3, 3, figsize=(20, 16))
flatax = ax.flatten()

for human, a in zip(test_humans, flatax):
    path = str(human) # ensure path is string
    
    # load image
    image = Image.open(path).resize(CROP)
    
    # run the app
    title = run_app(path)
    
    # display the answers
    a.imshow(image)
    a.set_title(title)
    a.set_axis_off()

plt.tight_layout()
In [33]:
# random number for starting index
dog_index = np.random.randint(100)

# select images
test_dogs = dog_files[dog_index:dog_index+9]

fig, ax = plt.subplots(3, 3, figsize=(20, 16))
flatax = ax.flatten()

for dog, a in zip(test_dogs, flatax):
    path = str(dog) # ensure path is string
    
    # load image
    image = Image.open(path).resize(CROP)
    
    # run the app
    title = run_app(path)
    
    # display the answers
    a.imshow(image)
    a.set_title(title)
    a.set_axis_off()

plt.tight_layout()
In [34]:
added_images = 'added_images/'
test_transform = transforms.Compose([
    transforms.Resize(RESIZE),
    transforms.CenterCrop(CROP),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

image_folder = datasets.ImageFolder(
   added_images, transform=test_transform
)

test_loader = torch.utils.data.DataLoader(
    image_folder, batch_size=5,
    shuffle=False, num_workers=0  
)

added_paths = np.array(glob("added_images/*/*"))
images, _ = next(iter(test_loader))
In [35]:
fig, ax = plt.subplots(2, 4, figsize=(26, 16))
flatax = ax.flatten()

for path, a in zip(added_paths, flatax):
    # ensure path is string
    path = str(path) 
    
    # load the images and 
    image = Image.open(path).resize(CROP)
    
    # run the app
    title = run_app(path)
    
    # display the answers
    a.imshow(image)
    a.set_title(title)
    a.set_axis_off()

plt.tight_layout()
In [ ]:
 
In [ ]: